How Nvidia’s Ampere architecture will improve the company’s gaming graphics cards remains a mystery. But on Thursday, Nvidia CEO Jensen Huang talked up how the technology will supercharge the company’s enterprise GPUs.
The Ampere architecture will first end up in the A100, a graphics card designed for data analytics and scientific computing. According to Nvidia, the A100 will offer a 20-times performance boost when running AI-powered applications compared with the company’s V100 GPU from three years ago.
The A100 GPU (Credit: Nvidia)
The new card itself contains 54.2 billion transistors, 6,912 CUDA Cores, a 40GB memory size, which can produce a memory bandwidth of 1.6 TB/sec, or what Huang said is an industry first. The company was able to do this with the help of TSMC’s 7-nanometer manufacturing technology, which can pack more transistors on the silicon. (Full specs for the A100 can be found here.)
In contrast, the older V100 GPU —which only has 21 billion transistors— is capable of a 900 GB/sec memory bandwidth using the older 12nm fabrication technology.
To illustrate the difference, Huang demoed the A100 running against the V100 over an AI application that lets you upload bird sounds to find which bird species and geographic region the audio comes from. The A100 was able to carry out 500 queries per second compared with only 80 queries per second offered by the V100.
Another advantage the A100 has over its predecessor is how the GPU’s computing power can be split down to as many as seven independent GPU tasks. This promises to help data centers and cloud providers drive down costs when leasing their computing power to clients. “Each one of the customers could rent a smaller computer. You now have the flexibility to do that,” Huang added.
Microsoft will be the first customer to adopt the A100 card, and plans on using the technology in the company’s cloud computing platform, Azure, which data scientists can use to fine-tune and run their AI programs. “Azure will enable training of dramatically bigger AI models using Nvidia’s new generation of A100 GPUs to push the state of the art on language, speech, vision and multi-modality,” said Microsoft corporate vice president Mikhail Parakhin in a statement.
The DGX A100 board (Credit: Nvidia)
The card is also arriving packaged as a prebuilt system with the DGX A100, a $199,000 server unit. Each unit comes with eight A100 GPUs slotted inside. The first DGX A100 order is going to the US Department of Energy’s Argonne National Laboratory to help virtually simulate COVID-19, which can help scientists develop potential treatments for the virus.
Although today’s talk by Huang focused little on the consumer side, the Ampere architecture is expected to bring a performance boost to Nvidia’s next-gen gaming graphics cards by ditching the 12-nm manufacturing tech for TSMC’s faster 7-nm fabrication process.
Expect Nvidia to talk more about Ampere in the coming months as the company prepares to release its rumored 3000-series RTX graphics cards for consumers.